GPU Chess Hackathon: Grandmasters aren’t born, they’re trained
Spend a weekend training AI Chess bots on GPUs.
TLDR:
This event is for people wanting to get their hands dirty in rapid iteration cluster training, using Strong Compute’s Instant Super Computer.
The main difference about how you will have trained before and this environment: The Instant Super Computer is a control system, storage and networking infra product that makes it much easier and faster to run jobs on shared GPU clusters.
Most of the time you should be able to start training within seconds, even for large datasets and large cluster sizes.
The same cluster can be shared by multiple people, with a lot of the headache for sharing resources taken care of.
Large datasets are available quickly and easily.
The goal is to experiment with different models, datasets and hyperparameters and train the best chess AI you can in a weekend.
The winner will receive a $10k-$100K compute grant for open-source AI research purposes.
Details:
We’re providing 2-day access to 48 GPU clusters. 10 teams will compete, blending human intuition, champion moves, and innovative techniques to create the ultimate chess AI.
Task: Train a language model to play chess. The input will be the game history in Portable Game Notation (PGN) using Standard Algebraic Notation (SAN). The model will generate characters until a move description is complete.
Dataset: Public dataset of training game PGNs from Leela Chess Zero (https://lczero.org/ - https://storage.lczero.org/files/training_pgns/test60/). An example data loading pipeline will be provided. You can use this or source your own.
Models: Must be character-level generative language models. Example models will be provided, or you can develop your own. Models must meet memory and inference time constraints to be determined.
Models must be trained from scratch during the competition weekend.
Compute: access to multiple 48x 24GB Ampere clusters and dedicated GPU workstations on the Strong Compute Instant Super Computer platform.
Budget: compute credits provided during the competition.
Judging: each team’s model will play off in a tournament to determine the competition winner.
Prize: winning teams will receive one of Strong Compute’s $10-100K grants for open source AI research - or further developing open chess models.
The full set of rules and gameplay will be shared with participants ahead of time on Discord.
Spots are limited. We’re prioritising participants with experience in PyTorch and cluster training.
ML training: you’ve run training jobs and know your way around a few GPUs.
Distributed training at scale: even better if you’re skilled in large scale clusters (64+ GPUs).
Teams: teams of 1 to 3 members are allowed. When you register, include your team members’ names. Approved participants will receive a link to our discord competition channel.
Location:
Artarmon, NSW, specific details to be provided to successful applicants.
Over discord for virtual participants.
Sat, 3 August:
9.00 am coffee, breakfast + networking
9.30 am set-up, onboarding, test training run + inference
10.30 am competition begins
12.30 pm lunch
6.00 pm dinner
12.00am wrap up day one
Sun, 4 August:
9.00am day two start, breakfast
12.30 pm lunch
2.00 pm tournament begins
3.30 pm winners announced
4.00 pm done
Strong Compute
Strong Compute provides infrastructure management capability for rapid artificial intelligence development.
P.S. We're hiring and this is a great way for us to get to know each other https://strongcompute.com/jobs
Supported by community partner:
Build Club: the home for top AI engineers in APAC